Skip to content

feat(ubenchx): combine B200 and Huawei core profiles / 合并 B200 与 Huawei 核心访存特征 - #1507

Merged
Oseltamivir merged 5 commits into
masterfrom
feat/ubenchx-ascend-core-profiles
Oct 26, 2026
Merged

Oseltamivir merged 5 commits into
masterfrom
feat/ubenchx-ascend-core-profiles

Conversation

@Oseltamivir

@Oseltamivir Oseltamivir commented Oct 26, 2026 •

Copy link
Copy Markdown
Contributor

Combines B200 and Huawei 950PR on one Core memory-access profiles page. Both appear by default; a hardware selector focuses on either device. The old B200 page redirects to the combined page in English and Chinese.

B200 retains its SM/GPC/die matrix in clock64 cycles. Huawei adds four 128 MiB Cube/Vector scalar-controller matrices, covering isolated and concurrent execution in raw SYS_CNT ticks. Each device keeps its own color scale and measurement conditions: these are matched-address profile differences, not direct core-to-core latency or physical coordinates. Concurrent Huawei launches lack a grid-wide start barrier.

The public API supports the same hardware selection and preserves the existing sm-l2-distance response. Cell inspection, pinning and clearing remain available. Merged latest master, including #1506 and #1508.

Validated both devices, hardware filtering, legacy redirects, Chinese rendering, API values and Huawei pin/clear behavior through Cypress. Static ubenchX captures do not use inference-run overlays. The broader local smoke run had one unrelated Chinese-blog math failure; an isolated recheck hit Next.js development instrumentation (UnknownChineseRoute negative performance timestamp). That failure is not suppressed; production CI remains the merge gate.

AI model disclosure
  • Primary implementation, data audit, tests and visual checks: the exact model identifier was not exposed by the Codex runtime and could not be verified.
  • claude-opus-5-5: advisory Chinese-copy review of the combined page and API documentation. The primary agent reviewed the suggestions and desktop/mobile renderings. An earlier review attempt for the initial PR was refused and contributed no feedback; the current review completed.
PR checklist
  • AI model disclosure is complete and current.
  • The combined view and hardware selector include matching API and skills documentation.
  • English and Chinese pages and legacy redirects are included.
  • Changed Chinese copy received advisory review and fidelity/rendering checks; final maintainer review remains applicable.
中文

将 B200 与 Huawei 950PR 合并到同一个核心访存特征页面,默认同时显示两者,也可通过硬件选择器单独查看。旧 B200 页面的中英文链接均重定向到合并后的页面。

B200 保留以 clock64 周期计的 SM/GPC/die 矩阵。Huawei 新增四幅 128 MiB Cube/Vector 标量控制器矩阵,分别覆盖单核执行和并发启动,单位为原始 SYS_CNT ticks。各设备保留独立色标和测试条件。这些数值是对应地址上的访存特征差异,不是核心间直接通信延迟或物理坐标;Huawei 并发启动没有网格级起始同步屏障。

公开 API 支持同样的硬件选择,旧 sm-l2-distance 响应结构保持兼容。单元格查看、固定和清除选择功能继续可用。已合并最新 master,包括 #1506 和 #1508。

已通过 Cypress 验证两种设备、硬件筛选、旧链接重定向、中文显示、API 数值及 Huawei 单元格固定与清除。静态 ubenchX 数据不使用推理运行叠加功能。完整本地 smoke 测试中,一项无关的中文博客数学渲染测试失败;单独复测触发 Next.js 开发模式性能计时错误(UnknownChineseRoute 时间戳为负)。未屏蔽该错误,合并仍以生产构建 CI 为准。

AI:主要实现、数据核对、测试和视觉检查所用模型的精确标识无法从 Codex 运行环境核实。claude-opus-5-5 完成了合并页面及 API 文档的中文辅助审核,主要 agent 核对了建议和桌面/移动端显示。初版 PR 的一次早期审核请求曾被服务拒绝,未产生反馈;本次审核已完成。

保留 Ascend Cube 与 Vector 标量访存矩阵、实际核心 ID、计时单位和来源哈希。
新增 Cube 与 Vector 的独立和并发访存特征页面,保留 SYS_CNT 单位与解释限制,并同步中英文页面、API 和技能文档。
@Oseltamivir
Oseltamivir requested a review from adibarra as a code owner October 26, 2026 12:44
@cursor

cursor Bot commented Oct 26, 2026

Copy link
Copy Markdown

Bugbot is paused — your team has an unpaid invoice

Bugbot is paused because your team has an unpaid invoice.

Pay it from the billing page in the Cursor dashboard and Bugbot resumes automatically.

(requestId: 28957b1c-f6c9-43c7-bec0-e506abea6a42)

@vercel

vercel Bot commented Oct 26, 2026 •

Copy link
Copy Markdown

The latest updates on your projects. Learn more about Vercel for GitHub.

Project Deployment Actions Updated
inferencemax-app Ready Ready Preview Oct 26, 2026 3:42pm UTC

Request Review

支持再次点击或按 Enter 取消固定,并通过图外点击、Escape 或清除按钮清空核心对选择。
@cursor

cursor Bot commented Oct 26, 2026

Copy link
Copy Markdown

Bugbot is paused — your team has an unpaid invoice

Bugbot is paused because your team has an unpaid invoice.

Pay it from the billing page in the Cursor dashboard and Bugbot resumes automatically.

(requestId: 03cf7bbc-a2a7-43e4-8834-8880e95a0146)

中文:将最新 master 合并到核心访存特征分支,包含 Huawei 带宽和统一配色。

# Conflicts:
#	packages/app/cypress/e2e/ubenchx.cy.ts
#	packages/app/src/app/api/v1/views/ubenchx/route.test.ts
#	packages/app/src/lib/api/api-route-catalog.ts
#	packages/app/src/lib/views-api/docs/extensions.ts
#	packages/skills/skills/inferencex-api/integrity.json
#	packages/skills/skills/inferencex-api/references/dashboard-views.md
中文:将 B200 与 Huawei 核心访存特征合并到同一页面,提供硬件筛选,保留独立单位与色标,并兼容旧页面链接和 API。
@Oseltamivir Oseltamivir changed the title feat(ubenchx): add Ascend core memory profiles / 新增 Ascend 核心访存特征 feat(ubenchx): combine B200 and Huawei core profiles / 合并 B200 与 Huawei 核心访存特征 Oct 26, 2026
@cursor

cursor Bot commented Oct 26, 2026

Copy link
Copy Markdown

Bugbot is paused — your team has an unpaid invoice

Bugbot is paused because your team has an unpaid invoice.

Pay it from the billing page in the Cursor dashboard and Bugbot resumes automatically.

(requestId: 3dcc0f96-a710-4dd1-96ce-5baffce2fdf3)

@Oseltamivir
Oseltamivir merged commit c108bf6 into master Oct 26, 2026
37 checks passed
@Oseltamivir
Oseltamivir deleted the feat/ubenchx-ascend-core-profiles branch October 26, 2026 16:11

This branch was successfully deployed

1 active deployment
Preview — bd805dc4 Deployed Oct 26, 2026 by vercel[bot]
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant